mg262
25th September 2005, 19:33
Here's a thought: suppose that we have adapted a global motion compensater to estimate zoom in x and y separately. Take a clip and find a scene that is a pure zoom (which can be done perfectly on animation and pretty well on a lot of other stuff). Now estimate horizontal and vertical zoom for a pair of frames, and you can recover the aspect ratio -- it's just the ratio of these two numbers. (Because in the original clip, horizontal and vertical zoom factors would have been the same).
You can improve the accuracy by processing more than two frames -- but even so, the number of frames that needs to be processed is quite small so there's quite a lot of computational power available. It may even be that we can do this using existing motion compensators and/or script functions using animate together with resizers. But because Fourier estimation methods essentially treat each dimension separately, I would guess that it wouldn't be hard to adapt them for this.
Something similar but more robust can be done with rotation. In fact, I think this can even be adapted to find the sampling kernel... but enough random thoughts for one post.
You can improve the accuracy by processing more than two frames -- but even so, the number of frames that needs to be processed is quite small so there's quite a lot of computational power available. It may even be that we can do this using existing motion compensators and/or script functions using animate together with resizers. But because Fourier estimation methods essentially treat each dimension separately, I would guess that it wouldn't be hard to adapt them for this.
Something similar but more robust can be done with rotation. In fact, I think this can even be adapted to find the sampling kernel... but enough random thoughts for one post.